Week to week attendance and competitive balance in the national football league.

Coates, Dennis ; Humphreys, Brad R.

Introduction

The relationship between competitive balance and attendance at sporting events has become an important area of research in sports economics over the past decade. Rottenberg (1956) first articulated the relationship between competitive balance and attendance, pointing out that attendance "is a negative function of the goodness of leisure-time substitutes for baseball in the area and of the dispersion of percentages of games won by the teams in the league" (p. 246). Rottenberg's conjecture that attendance demand depends in part on the dispersion of winning percentages, a common measure of competitive balance in sports leagues, is now known as the Uncertainty of Outcome Hypothesis (UOH) in the sports economics literature. Interest in the relationship between competitive balance and attendance increased early in the last decade when Zimbalist (2002) suggested that the competitive balance-attendance relationship should be the focus of competitive balance research, as understanding fans' perceptions of competitive balance (as reflected in the sensitivity of attendance to changes is different measures of competitive balance) was a key way to evaluate the usefulness of the ever expanding catalog of ways to measure competitive balance. Humphreys and Watanabe (in press) survey the literature on the UOH and the analysis of competitive balance following Zimbalist's (2002) special issue of the Journal of Sports Economics on competitive balance.

Early research on the UOH focused on relatively aggregated measures of demand like total season attendance in sports leagues, or total season attendance at home games played by teams in sports leagues. Borland and Macdonald (2003) survey the early literature on attendance demand and uncertainty of outcome. Two important developments recently emerged in this literature: the use of temporally disaggregated data, typically game or match level data, instead of season total attendance at the league or team level, and the use of measures of uncertainty of outcome that extend beyond the simple measures of dispersion of winning percentages posited by Rottenberg (1956).

This paper contributes to the growing literature on the relationship between attendance and competitive balance using match or game level data and a variety of UOH measures and game quality. We investigate the determinants of live game attendance at National Football League (NFL) games, including factors related to uncertainty of outcome as well as current and past team quality and game day characteristics. Little research has focused on game attendance at NFL games, or on the effect of outcome uncertainty on attendance at NFL games. This lack of research is somewhat surprising, given the immense popularity of the NFL.

Welki and Zlatroper (1999) estimated a demand function for game day attendance at NFL games in the 1986 and 1987 seasons and found that team quality as measured by the teams' season winning percentages prior to the game, and uncertainty of outcome, as measured by the point spread, among other factors, explained observed game day attendance. Carney and Fenn (2004) analyzed the determinants of the television viewing audience for regular season NFL games in the 2000 and 2001 seasons and found that team quality as measured by the teams' season winning percentages prior to the game and the closeness of the game as measured by the actual score difference in the game, among other factors, affected the size of the television viewing audience. The actual score difference is not a measure of uncertainty of outcome, as it is known with certainty by the econometrician. Paul and Weinbach (2007) analyzed the determinants of the television viewing audience at the beginning of the game and at halftime for regular season NFL games over the period 1992-2002. They found that the quality of the teams, as measured by the sum of the teams' winning percentages prior to the game, the expected closeness of the contest, as measured by the difference in the teams' winning percentages prior to the game, and the expected offense in the game, as measured by the over/under betting line affected the size of the television audience at beginning of the game, and that scoring in the first half affected the size of the second half television audience. Alavy et al. (2010) assess the impact of uncertainty of outcome on television viewership in English football, using minute-by-minute television ratings as their outcome. They find that viewership drops off as a game looks more and more like it will be a draw and that games that end with victories have higher viewership on average than games that end in ties.

Biner (2009) analyzed the determinants of both television ratings (for 491 of the 1943 NFL games played in the 1972-1978, 1981, and 1983 seasons) and game day attendance for the 1994-2007 NFL seasons using reduced form parametric and semi parametric regression models. He found that the quality of the home team, as measured by the home team's winning percentages prior to the game but not the visiting team, was associated with higher game day attendance and that uncertainty of outcome, as measured by the absolute value of the point spread, was associated with lower game day attendance. While Biner (2009) also examines outcome uncertainty in the NFL, that paper focuses on television ratings while we focus on the effect of past success on live game attendance.

We extend this literature by analyzing the determinants of game day attendance at more than 5,000 regular season NFL games played in the 1985 through 2008 seasons, focusing on the role played by uncertainty of outcome, quality of teams, game characteristics, and fan expectations based on previous performance by the home team in the current and past season and of the visiting team in the current season. Our data includes more games from more seasons than either Biner (2009) or Welki and Zlatroper (1999) as well as more franchise and game specific determinants of attendance variables than either. The main source of data is the NFL website and a now (apparently) defunct site that had box scores that included game day attendance. Our data set encompasses all the seasons analyzed in previous papers and a larger array of explanatory variables. Moreover, we use our estimates to ask a unique question about the appropriate level of competitive balance in the NFL. Specifically, if there was complete balance, would attendance be higher for some or all teams?

We find weak evidence that fans prefer to see dominant home team wins, in that the estimated coefficient on the absolute value of the point spread on the game is positive and weakly significant. We also find that fans are not interested in watching the home team lose games, in that the estimated parameter on the point spread variable when the home team is an underdog, is negative and significant. In other words, our results do not support the traditional interpretation of the uncertainty of outcome hypothesis that attendance would be greater at games in which clubs are more evenly matched. We also find that the effect the previous season success has on attendance in the current season is completely removed by the inclusion of average attendance from the previous season.

Empirical Analysis

Data Description

Data for this analysis comes from 5,535 regular season National Football League games held during the seasons from 1985 through 2008 though the preferred regression specification includes only 5,270. The data were obtained from the NFL website. During this time the league organization changed from two conferences with three divisions each to two conferences with four divisions each. Several teams relocated and a number of expansion teams came into the league. The season was also extended and a bye week introduced. The NFL follows a scheduling procedure in which clubs' fixtures for this season are dependent on the previous season's outcomes. Successful teams from last season have more games against other such teams on their schedule this year while less successful teams from the previous year play one another more often this year. This scheduling format is a clear signal that the NFL believes that more evenly matched games and greater uncertainty of outcome are beneficial to its bottom line. Our treatment of relocation and some expansion deserves more discussion. Consider that during our sample period the Cardinals left St. Louis for Phoenix, the Raiders left Los Angeles to return to their original home of Oakland, the Los Angeles Rams went to St. Louis, the Browns left Cleveland for a new home in Baltimore as the Ravens, and the Oilers left Houston to become the Tennessee Titans. Cleveland and Houston also regained teams. Jacksonville and Carolina joined the NFL as pure expansion franchises where no NFL team had previously existed. In our analysis, when cities have only one franchise, one can think of the city as the observation. For example, home games in St. Louis are not distinguished by whether the club is the Cardinals or the Rams; home games in Houston are treated the same whether they are Oilers or Texans games. Similarly, the Oakland Raiders are treated as a distinct entity from the Los Angeles Raiders and the Los Angeles Rams and the St. Louis Rams are different. The Tennessee Titans are treated as distinct from the Houston Oilers and the original Cleveland Browns are distinct from the Baltimore Ravens.

Table 1 reports descriptive statistics for the variables in our analysis. The dependent variable is the natural logarithm of (home) game attendance for games involving home team i against visiting team j in week t of season s.

Empirical Model

We formulate a reduced form regression model for the determination of game day attendance at NFL games. The model includes explanatory variables that have been identified as important determinants of game day attendance in this literature. This empirical model can be motivated by a model of profit or win maximizing teams and utility maximizing consumers. One key factor identified in the literature as a determinant of game day attendance is uncertainty of outcome. Winning percentages of the home and visiting teams are one possible way to capture uncertainty of outcome. But winning percentages also capture the quality of the teams. We also use the point spread on the game as a proxy for uncertainty of outcome. The point spread ranges from negative 24 to positive 23, indicating a very strong home favorite to a very extreme home underdog. The larger in absolute value is the point spread the less competitive the game is expected to be. Low values of the absolute value of the point spread indicate more competitive games, which may be of more interest to fans and, consequently attract greater attendance. Additionally, to assess the possibility that attendance responds differently to home team favorites and home team underdogs, we create a home underdog dummy variable which we interact with the absolute value of the point spread. This construction allows for underdog status and favorite status to affect attendance asymmetrically, something the quadratic form of Welki and Zlatoper (1999) does not allow. Finally, we include an interaction between the absolute point spread and a dummy variable indicating a game against a divisional rival.

Variables used to explain attendance at each game include the previous season's average home game attendance, the winning percentage of the home and visiting teams in all games that season prior to this game, the points scored and points allowed per game by the home and visiting team to that point in the season, and the home team's winning percentage from the previous season. The analysis also includes dummy variables indicating games played in the first week of the season, (1) whether the teams were from the same division, whether they were from the same conference, if the stadium has a dome or roof, and if the field is artificial turf. We also utilize a variable that counts the week of the season, from 1 to 17 and dummy variables indicating games played in October, in November, or in December.

Finally, we include variables intended to measure the extent to which the influence of last season's success, or lack of success, on the field evaporates as the current season passes. Our interest is in how quickly last season's result is forgotten, replaced by the franchise's current season competitiveness in determining attendance. The variables to capture this effect could take a variety of forms. We include the previous season's winning percentage interacted with the week of the season variable. When the dependent variable in the analysis is the natural log of attendance, the coefficient on this interaction term is the rate of decay in attendance. To see this, let attendance at the game between home team i and visiting team j in week t of season's be given by

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

Taking the natural log of attendance produces our basic empirical model

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

The vector of explanatory variables, [X.sub.ijtsk], includes home and visiting team winning percent to date, previous season average home game attendance, and so on. [theta] is the estimated rate of decay of the previous season's winning percent on the current season's attendance.

That specification of the log of attendance is quite general, and is the approach we take in our estimation. The equation error, [v.sub.ijts], has a home team and a visiting team specific component and a purely random component. That is, [v.sub.ijts] = [[mu].sub.i] + [[omega].sub.j] + [z.sub.ijts]. Of course, attendance is constrained by stadium capacity. Consequently, our empirical approach is to estimate the model using a Tobit estimation technique with regression errors clustered by home team. Following Buraimo and Simmons (2008), we define attendance as constrained when it is at or above 95% of stadium capacity.

Results

Table 2 reports estimation results for two models. Model 1 omits average attendance from the previous season and the square of the absolute point spread. In all other respects the two models are identical. The results are largely in line with intuition. There is a large blip in attendance, about 20%, associated with the opening game of the season. Games against divisional rivals get 7% more attendance than non-divisional games, though the larger the absolute point spread, the smaller this divisional opponent effect.

Interestingly, the evidence here indicates that high scoring home teams do not attract greater attendance, though high scoring visitors do. Additionally, home teams that give up lots of points experience slightly reduced attendance in Model 2. The effect is about 2% less attendance for each five points per game extra a team allows. Five points per game is only about 70% of a standard deviation in the points allowed per game variable, so it isn't a small effect. Moreover, because we control for the team's winning percentage, this points allowed affect is not simply that losing teams draw poorly; instead this result suggests that defense not only wins championships but it also attracts fans.

The month dummy variables assess the importance of weather and other seasonal effects. The results clearly reveal that as the season progresses attendance declines. For example, games in December attract about 12% lower attendance than games in September. October games draw fewer fans than September games, though games in November do not. Note however that December games played in domed stadiums or in the sunbelt attract about 8 to 9% more attendance. Monday night games are a big draw, adding about 12% to attendance, but Thursday, Friday and Saturday games do not have a similarly beneficial impact on attendance. New stadiums, meaning those in their first year of operations, generate a substantial increment to attendance on the order 20%.

The remaining variables in the analysis are related to the importance of competitive balance and the uncertainty of outcome hypothesis. The team's winning percentage from the previous season can be an indicator of team quality and a predictor of its success this season. Fans attracted to games because of the quality of a team would likely update their perception of quality as new evidence is provided and the influence of last season's winning percentage on this season's attendance is expected to decline as this season unfolds. The winning percentages of the two teams up to this point in the season and the betting point spread are two indicators of the uncertainty of the outcome.

The results reported in Table 2 Model 1 indicate that the better a team was last season, the greater its attendance this season. However, once the average attendance from the previous season is controlled for in Model 2, the influence of the previous season success disappears. Indeed, our attempt to estimate the rate of decay of previous season success on current season attendance is a failure. None of the variables intended to capture this decay are individually significant nor are they jointly significant. Moreover, as noted, even previous season success is insignificant once the model includes average attendance from the previous season. This suggests that a significant impact of previous season winning percentage may be a proxy for the habit or persistence of fans to attend games of their team. The coefficients on the home and visiting team current season winning percentage to date variables are positive and statistically significant. The home team variable has twice to three times the coefficient as the visiting team variable, however. This result indicates that fans want to see good teams play, even if one of them is the visiting team, but a quality home team is still more important than the quality of the visitor. On the other hand, the point spread variables indicate that home teams that are big favorites draw better than home teams expected to be in a tight contest and far better than home underdogs. In fact, a one-point increase in the spread for a home favorite raises attendance by about seven-tenths of a percent in Model 1 and about 1.4 percentage points in Model 2, while a one-point increase in the spread against a home underdog reduces attendance by about one percentage point in both specifications.

These results can be summed up simply. Consistently good teams get a benefit of the doubt from their fans while consistently bad teams do not. Fans want to see good teams play football. They want to see the home team win big, but they are not as interested in watching the home team lose big, holding the quality of the teams constant.

Our empirical model allows us to estimate the effect of the visiting opponent on home attendance. These reduced form parameters capture the net effect of all factors related to the opposing team on home attendance, including any fans of the visiting team that have travelled to attend the game. Table 3 shows the visiting team effects, with the Washington Redskins as the omitted category. From the perspective of understanding the economic impact of professional sports on the local economy, the key question is whether some teams travel well, in the sense that large numbers of their fans travel to away games. In the debate over stadium subsidies an important issue is how many of the fans attending a game have travelled from out of town to see the game. The larger the number of outside visitors, the better the chance that games generate net increases in hotel stays, meals in restaurants, and so on, and therefore generate a net gain to the community. The first column indicates the coefficient on the visiting team indicator variables from Model 1. This estimated parameter captures the net effect of all factors associated with the visiting team on home attendance, and can include factors like the number of fans the visiting team has in the home team's market, the reputation of the visiting team, and potentially the propensity of fans of the visiting team to travel to away games. Given the size of these parameter estimates, most may be reasonably close approximations to the proportionate contribution to attendance of each visiting franchise. The parameter estimates indicate that most teams have a smaller number of fans that travel to away games than do the Redskins. Only the Dallas Cowboys have statistically significant and positive parameter estimates, indicating attendance rises when they are the visiting team even more than it does when the Redskins are the visitors. To the extent that these parameter estimates reflect the number of visiting team fans that travel to away games, the results do not indicate that a large number of fans engage in this type of travel. This result suggests that any economic impact associated with fan spending on hotels, bars, restaurants, and other attractions by visiting team fans traveling to the home team's market is small.

Table 4 reports the home team effects. In other words, after controlling for other influences, these coefficients indicate which teams draw well at home relative to the Washington Redskins. The results here suggest two things. First, warm weather teams generally draw relatively poorer than do cold weather teams, all else constant. Arizona, Carolina, and the Florida- and California-based franchises all have negative and statistically significant coefficients in Model 2. By contrast, Baltimore, Cleveland, Green Bay, and Philadelphia have positive and significant coefficients. Second, two cities that lost, and then regained, franchises, Baltimore and Cleveland, are in the top three in terms of home team boost to attendance, while cities/regions that got franchises from relocations or expansion, such as Los Angeles (twice), Arizona, Jacksonville, Carolina, Indianapolis, and St. Louis only have average home drawing power or even have negative and statistically significant home effects.

Forrest et al. (2005) conduct an interesting analysis of the impact that improved competitive balance would have on attendance. A variety of league rules within professional sports in the United States, especially, are defended or proposed as means of enhancing competitive balance. Among these are the reverse-order draft, revenue sharing, salary caps, and luxury taxes. Leaving aside the issue of whether these institutions have been or are likely to be successful at improving competitive balance, we can, like Forrest et al. (2005), ask the question of how improved competitive balance would affect attendance. For the English Football League, based on 844 matches from the 1997-98 season and extrapolated to the full season, they found that equality of playing ability across clubs would reduce aggregate attendance by over 2 million. Our results are not quite so dramatic, but large nonetheless.

Using model 2 we calculated the predicted attendance if every team won half of its games, scored the same number of points each game, had the same previous season winning percentage, the point spread was zero, and attendance could not exceed capacity. The average of this predicted game attendance is 55,678, compared to the actual average attendance of 62,465, a difference of 6,787 fewer attendees per game. This amounts to about an 11% drop in attendance, a total of over 35.77 million fewer attendees over the 22 years of the sample or about 1.63 million per year. Predicted attendance is lower than actual attendance in 76.5% of the games in the sample. For comparison purposes, we also calculate predicted attendance at the observed values of the explanatory variables, given the censoring and uncensored, and under perfect balance assuming no censoring. Mean values of these variables, respectively, are 56,477, 75,741, and 74,176. These results indicate that stadium capacities significantly constrain attendance at NFL games but that perfect competitive balance would harm attendance.

Because our analysis covers multiple seasons, we can examine the difference from one season to the next in attendance under perfect balance compared to actual balance. Doing this we see that predicted attendance under perfect balance (and attendance limited by capacity) is lower than actual attendance, on average, in every year in our sample. The implication is that actual competitive balance in those years resulted in greater attendance than would have occurred under perfect balance.

An additional issue we can address is the extent to which individual clubs will benefit or be harmed by greater balance. Comparing the actual per game attendance to the predicted attendance under complete equality of teams and censoring by stadium capacity, on a team by team basis, reveals that every team except the Arizona Cardinals would have had significantly lower attendance under perfect competitive balance relative to their actual attendance. The Cardinals' attendance would have been the same with a perfectly balanced league as it actually was. In other words, no team's attendance would have been improved had the NFL been completely balanced over the period 1986 through 2008, the time frame of this study.

Alternatively, suppose we compared predicted attendance without stadium capacity constraints, under either actual competitive balance or under perfect balance, against actual attendance. In the former case, all but the LA franchises, Raiders and Rams, would have seen statistically significantly more attendees had the stadiums had larger capacity. (2) Said differently, under the actual competitive balance within the NFL, every club except the Los Angeles Raiders and the Los Angeles Rams would be able to sell more tickets if stadium size were larger. These two clubs, which no longer exist in Los Angeles, are the only ones for which stadium capacity was not a binding constraint on attendance. Under perfect competitive balance, the Los Angeles Rams' attendance without a capacity constraint would have been significantly lower than its actual level. For both the LA Raiders and the Jacksonville Jaguars, perfect balance with no capacity constraint would not result in significantly more attendance than they actually had. For San Diego, the unconstrained attendance under perfect balance would have exceeded actual attendance, but only at the 10% level of statistical significance.

Discussion

Our results indicate that uncertainty of outcome does not play a large role in determining game day attendance at NFL games. Fans do not turn out for games expected to be close contests. Instead, fans (weakly) prefer to attend games that their team should win, and avoid games that the home team is expected to lose, holding the quality of the home team and the opponent constant. This result has important implications for sports league policy. Sports leagues rationalize policies like salary caps, reverse entry drafts, and limits on free agency by claiming that these policies must be put into place to promote competitive balance. If competitive balance declines, and some teams win a disproportionate number of games, leagues argue, then overall league revenues will suffer.

Our results suggest that fans like to see games in which they expect the home team will win by a large margin. Fans do not buy tickets to see close games, or games that they expect the home team to lose. However, there are many different types of uncertainty of outcome beyond the game uncertainty we examine here. Season uncertainty and championship uncertainty have also been identified as important components of the UOH. Home teams could win every home game by a large margin and the league championship could still be closely decided. Such an outcome would maximize home attendance and still provide a high degree of competitive balance in the league. League policies designed to promote competitive balance operate primarily at the season level, or over multiple seasons, and not at the game level. Fans' preferences for game uncertainty differ from league policy objectives, in that fans appear to prefer unbalanced game outcomes and leagues claim to prefer balanced outcomes over a season or seasons. Resolving this tension between fans and leagues appears to be an important topic for future research.

Finally, our empirical model does not include a price variable. Since economic theory predicts that prices are an important determinant of demand for live game attendance at sporting events, our results may be influenced by omitted variables bias. This could be especially important in the NFL, where most teams operate at or near capacity in most games in a season. Profit maximizing teams can price tickets in a way to guarantee a sellout, or near sellout in every game, leading to a systematic relationship between the omitted price variable and several of our explanatory variables, including previous season success.

References

Alavy, K., Gaskell, A., Leach, S., & Szymanski, S. (2010). On the edge of your seat: Demand for football on television and the uncertainty of outcome hypothesis. International Journal of Sport Finance, 5, 75-95.

Amemiya, T. (1984). Tobit models: A survey. Journal of Econometrics, 24(1-2), 3-61

Biner, B. (2009). Equal strength or dominant teams: Policy analysis of NFL. MPRA Paper 17920, University Library of Munich, Germany.

Borland, J., & Macdonald, R. D. (2003). Demand for sport. Oxford Review of Economic Policy, 19(4), 478-502.

Buraimo, B., & Simmons, R. (2008). Do sports fans really value uncertainty of outcome?: Evidence from the English Premier League. International Journal of Sport Finance, 3, 146-155.

Carney, S., & Fenn, A. J. (2004, October). The determinants of NFL viewership: Evidence from Nielsen ratings. Colorado College Economics and Business Working Paper No. 2004-02. Retrieved from http://ssrn.com/abstract=611721

Forrest, D., Beaumont, J., Goddard, J., & Simmons, R. (2005). Home advantage and the debate about competitive balance in professional sports leagues. Journal of Sports Sciences, 23(4), 439-445.

Humphreys, B. R. (2002). Alternative measures of competitive balance in sports leagues. Journal of Sports Economics, 3(2), 133-148.

Humphreys, B. R., & Watanabe, N. (in press). Competitive balance. In L. Kahane & S. Shmanske (Eds.), The Oxford handbook of sports economics, Volumes I and II. Oxford University Press: New York.

Paul, R. J., & Weinbach, A. P. (2007). The uncertainty of outcome and scoring effects on Nielsen ratings for Monday Night Football. Journal of Economics and Business, 59(3), 199-211.

Rottenberg, S. (1956). The baseball players' labor market. Journal of Political Economy, 64, 242-258.

Spenner, E. L., Fenn, A. J., & Crooker, J. (2004, October). The demand for NFL attendance: A rational addiction model. Colorado College Economics and Business Working Paper No. 2004-01. Retrieve from http://ssrn.com/abstract=611661

Welki, A., & Zlatoper, T. (1999). U.S. professional football game-day attendance. Atlantic Economic Journal, 27(3), 285-298.

Zimbalist, A. S. (2002). Competitive balance in sports leagues: An introduction. Journal of Sports Economics, 3, 111-121.

Endnotes

(1) We also included a dummy variable indicating a team's home opener. This variable was not statistically significant. In the data, a small number of cases occurred where the home opener was in the fourth or fifth week of the season.

(2) It should be noted that these clubs played in the Los Angeles Memorial Coliseum which hosted games with announced attendance over 100,000. One such game, between the Rams and San Francisco 49ers in 1957, set the one-game paid attendance record of 102,368, which stood until broken in September 2009 in the then newly opened Dallas Cowboys stadium.

Authors' Note

Brad Humphreys thanks the Alberta Gaming Research Institute for funding that supported this research.

Dennis Coates [1] and Brad R. Humphreys [2]

[1] University of Maryland, Baltimore County

[2] University of Alberta

Dennis Coates is a professor in the Department of Economics. His research interests focus on the effects of stadiums and professional sports on local economies.

Brad R. Humphreys is a professor in the Department of Economics and chair in the economics of gaming. His current research focuses on the economic impact of professional sports and the economics of sports gambling.

Table 1: Descriptive Statistics

Variable Mean Std. Dev. Min Max

Log of Game Attendance 11.019 0.239 7.910 11.547

Lagged Average Attendance 11.020 0.179 10.074 11.403
(log; previous season)

Home team winning percent 46.206 28.759 0 100
to date

Visiting team winning 47.743 28.845 0 100
percent to date

First week of season 0.063 0 1

Home Points scored per 19.056 7.391 0 45
game to date

Home Points allowed per 19.262 7.256 0 52
game to date

Visitor Points scored per 19.423 7.601 0 52
game to date
Visitor Points allowed per 19.097 7.237 0 51
game to date
Absolute Point Spread 5.507 3.494 0 24

Absolute Point Spread 42.537 54.234 0 576
Squared
Absolute Point Spread 1.442 2.697 0 23
- home underdog
Domed stadium 0.193 0 1

Artificial turf stadium 0.264 0 1

Teams from same division 0.387 0 1

Teams from same conference 0.759 0 1

Previous season winning 0.501 0 1
percent

Home winning pct. to 24.558 19.661 0 100
date*Prev. season win pct.

Week of the season*Prev. 4.537 3.145 0 16
season winning pct.

Week of the season 9.041 4.921 0 117

October 0.247 0 1

November 0.263 0 1

December 0.255 0 1

Monday 0.07 0 1

Thursday 0.017 0 1

Friday 0.004 0 1

Saturday 0.033 0 1

Dome*December 0.049 0 1

Sunbelt*December 0.063 0 1

Sunbelt 0.227 0 1

Sunbelt*September 0.049 0 1

Katrina 0.003 0 1

Expansion 0.003 0 1

New Stadium 0.024 0 1

Observations 5270

Table 2: Log Attendance Regressions

 Model 1 Model 2

 Coefficient p-value Coefficient p-value

Lagged Average 0.7478 *** 0.000
Attendance

Home team winning 0.0033 *** 0.000 0.0019 ** 0.021
percent to date

Visiting team 0.0011 *** 0.000 0.0009 *** 0.000
winning percent
to date
First week of 0.2067 *** 0.001 0.1546 *** 0.003
season

Home Points scored 0.0019 0.297 0.0028 0.165
per game to date

Home Points allowed -0.0029 0.199 -0.0045 * 0.053
per game to date

Visitor Points 0.0039 *** 0.003 0.0042 *** 0.001
scored per game
to date

Visitor Points -0.001 0.373 -0.0012 0.268
allowed per game
to date

Absolute Point 0.0074 ** 0.014 0.0143 ** 0.029
Spread

Absolute Point -0.0005 0.193
Spread Squared

Absolute Point -0.0108 *** 0.009 -0.0104 *** 0.005
Spread*home underdog

Domed stadium -0.0916 0.435 -0.0752 0.111

Artificial turf -0.0948 0.309 -0.0222 0.712
stadium

Absolute Point -0.0077 ** 0.029 -0.0080 ** 0.015
Spread*Teams from
same division

Teams from same 0.0778 *** 0.002 0.0780 *** 0.001
division

Teams from same -0.0211 * 0.053 -0.0151 0.153
conference

Previous season 0.3398 *** 0.003 0.0297 0.775
winning percent

Home winning pct. -0.0028 0.15 -0.0013 0.428
to date*Prev.
season win pct.

Week of the season 0.0036 0.607 0.0066 0.325
*Prev. season

Week of the season -0.0036 0.366 -0.0053 0.183

October -0.0305 * 0.093 -0.0282 0.125

November -0.0224 0.371 -0.0124 0.625

December -0.1241 *** 0.001 -0.1104 *** 0.002

Monday 0.1437 *** 0.000 0.1247 *** 0.000

Thursday 0.1156 0.121 0.1144 0.143

Friday -0.0933 0.247 -0.0985 0.223

Saturday -0.0111 0.739 -0.0227 0.479

Dome*December 0.0930 *** 0.001 0.0856 *** 0.001

Sunbelt*December 0.0749 *** 0.004 0.0836 *** 0.001

Sunbelt -0.2100 *** 0.000 0.0536 0.230

Sunbelt*September 0.0012 0.963 0.0091 0.739

Katrina -0.1164 * 0.053 0.0149 0.769

Expansion -0.0924 0.681 -0.0892 0.676

New Stadium 0.1894 *** 0.002 0.2190 *** 0.000

Constant 11.2165 *** 0.000 2.9883 *** 0.000

 0.2701 *** 0.000 0.2462 *** 0.000

Observations 5,535 5,270
Psuedo R-sq 0.539 0.628

Table 3: Visiting Team Effects

 Model 1 Model 2

 Coefficient t-stat Coefficient t-stat

Arizona -0.139 -4.12 -0.129 -3.76
Atlanta -0.121 -3.43 -0.103 -3.00
Baltimore -0.074 -2.28 -0.063 -1.95
Buffalo -0.078 -2.70 -0.068 -2.48
Carolina -0.082 -2.07 -0.080 -2.37
Chicago 0.066 1.22 0.068 1.18
Cincinnati -0.124 -3.70 -0.105 -3.23
Cleveland -0.059 -1.13 -0.053 -1.09
Dallas 0.137 2.30 0.145 2.35
Denver 0.000 0.02 -0.002 -0.08
Detroit -0.081 -2.12 -0.072 -1.88
Green Bay -0.017 -0.47 -0.012 -0.33
Houston -0.150 -4.06 -0.135 -4.04
Indianapolis -0.104 -2.55 -0.096 -2.61
Jacksonville -0.174 -4.19 -0.147 -3.70
Kansas City -0.063 -1.73 -0.052 -1.44
Los Angeles Raiders 0.078 1.52 0.079 1.64
Los Angeles Rams -0.052 -1.36 -0.038 -0.97
Miami 0.005 0.11 0.009 0.24
Minnesota -0.055 -1.60 -0.037 -1.08
New England -0.076 -2.12 -0.070 -2.14
New Orleans -0.113 -2.97 -0.108 -3.12
New York Giants -0.043 -1.35 -0.029 -0.86
New York Jets -0.088 -2.62 -0.077 -2.39
Oakland -0.065 -1.93 -0.040 -1.18
Philadelphia -0.072 -2.46 -0.075 -2.77
Pittsburgh 0.040 1.12 0.042 1.03
San Diego -0.135 -4.65 -0.126 -4.34
Seattle -0.102 -2.64 -0.098 -2.62
San Francisco 0.045 0.91 0.049 1.05
St. Louis -0.117 -2.61 -0.101 -2.11
Tampa Bay -0.150 -3.90 -0.127 -3.39
Tennessee -0.165 -3.82 -0.128 -3.52

Washington Redskins is the omitted category.

Table 4: Home Team Effects

 Model 1 Model 2

 Coefficient t-stat Coefficient t-stat

Arizona -0.181 -5.89 -0.250 -6.01
Atlanta -0.201 -2.99 -0.051 -1.51
Baltimore 0.303 3.85 0.266 8.28
Buffalo 0.108 1.17 0.007 0.13
Carolina -0.105 -3.89 -0.086 -4.89
Chicago -0.168 -13.81 -0.109 -8.30
Cincinnati -0.011 -0.12 -0.004 -0.06
Cleveland 0.289 14.13 0.172 7.28
Dallas 0.132 1.42 0.080 1.33
Denver 0.180 11.73 0.071 4.22
Detroit -0.088 -0.76 -0.084 -1.80
Green Bay -0.016 -1.08 0.053 4.48
Houston

Indianapolis 0.111 1.17 0.054 1.03
Jacksonville 0.01 0.09 -0.296 -5.76
Kansas City -0.028 -0.58 0.079 2.54
Los Angeles Raiders 0.112 2.33 -0.219 -4.61
Los Angeles Rams 0.013 1.39 -0.207 -4.3
Miami 0.18 8.84 -0.125 -2.64
Minnesota 0.024 0.22 0.052 1.09
New England 0.014 0.17 0.056 1.05
New Orleans 0.178 1.47 -0.061 -3.61
New York Giants 0.264 2.93 0.094 1.57
New York Jets 0.119 1.49 0.015 0.30
Oakland -0.306 -11.98 -0.154 -6.82
Philadelphia 0.283 3.00 0.185 2.98
Pittsburgh -0.031 -0.43 0.000 0.01
San Diego -0.080 -3.11 -0.254 -5.45
Seattle -0.037 -0.42 -0.014 -0.39
San Francisco -0.084 -3.67 -0.051 -2.78
St. Louis -0.079 -0.92 0.022 0.56
Tampa Bay 0.076 3.60 -0.083 -1.98
Tennessee -0.074 -0.89 0.300 3.62

Washington Redskins is the omitted category.